Trending Now - Deepseek Launches Affordable V4 Pro and V4 Flash AI Models

WHO:

Engadget - DeepSeek claims its V4 models have "world-class" reasoning capabilities.

South China Morning Post - DeepSeek, a Hangzhou-based AI start-up, released its new models.

South China Morning Post - Huawei announced "full support" for DeepSeek's V4 models.

South China Morning Post - Cambricon Technologies announced compatibility with DeepSeek’s new models.

WHAT:

AP News - DeepSeek's V4 Pro Max version has "superior performance" in standard reasoning benchmarks relative to OpenAI’s GPT-5.2 model and Google’s Gemini 3.0-Pro, but falls "marginally" short of GPT-5.4 and Gemini 3.1-Pro.

AP News - The V4 “pro” version could outperform Claude’s Sonnet 4.5 and approaches the level of Claude’s Opus 4.5 model based on DeepSeek's evaluation.

AP News - DeepSeek's V4 models are supported in part by computer chips made by Huawei, reducing reliance on U.S. chipmakers like Nvidia.

Techmeme - DeepSeek's V4 Pro model is priced at $1.74 per 1 million input tokens and $3.48 per 1 million output tokens, while the V4 Flash model costs $0.14 per 1 million input tokens and $0.28 per 1 million output tokens.

Techmeme - DeepSeek's V4 Pro model has a total of 1.6 trillion parameters, while the V4 Flash model has 284 billion parameters.

South China Morning Post - DeepSeek's V4 models have achieved a context window of 1 million tokens, a significant increase from the previous model's context window of 128,000 tokens.

VentureBeat - DeepSeek-V4-Pro is priced at approximately $5.22 for a simple one-million-input, one-million-output comparison, while DeepSeek-V4-Flash costs a combined $0.42.

VentureBeat - DeepSeek-V4-Pro costs about one-seventh the cost of GPT-5.5 and about one-sixth the cost of Claude Opus 4.7.

VentureBeat - DeepSeek-V4-Flash is priced at $0.14 per million input tokens and $0.28 per million output tokens, making it more than 98% below GPT-5.5 and Claude Opus 4.7 in a simple input-plus-output comparison.

MarkTechPost - DeepSeek-V4-Pro has 49 billion activated parameters per token, while DeepSeek-V4-Flash has 13 billion activated parameters per token.

South China Morning Post - DeepSeek's V4 models are described as "cost-effective" with a context length of 1 million tokens.

TechCrunch - DeepSeek states that both V4 models have "closed the gap" with current leading models on reasoning benchmarks.

Techmeme - DeepSeek's V4 Pro is the company's largest model by parameter count, featuring 1.6 trillion parameters, while the V4 Flash model has 284 billion parameters.

South China Morning Post - DeepSeek's V4 models have achieved a context window of 1 million tokens, a significant increase from the previous model's context window of 128,000 tokens.

South China Morning Post - DeepSeek claims that its V4 models are more efficient and performant than the previous V3.2 model due to architectural improvements.

Engadget - DeepSeek claims that the V4 Pro model has enhanced agentic capabilities and rivals top closed-source models in reasoning.

Engadget - DeepSeek states that the V4 Pro trails only Gemini-3.1-Pro in rich world knowledge.

Engadget - The V4 Flash model has faster response times compared to the V4 Pro and performs on par with the Pro version on simple agent tasks.

South China Morning Post - DeepSeek released two versions of its V4 model: V4 Pro and V4 Flash.

Techmeme - The V4 Pro model costs $1.74 per 1 million input tokens and $3.48 per 1 million output tokens, while the V4 Flash model costs $0.14 per 1 million input tokens and $0.28 per 1 million output tokens.

South China Morning Post - The V4 Pro model has 1.6 trillion parameters, and the V4 Flash model has 284 billion parameters.

Bloomberg - DeepSeek stated that the V4 Pro trails the performance of state-of-the-art models by about 3 to 6 months.

WHERE:

South China Morning Post - The release took place in Hangzhou, China.

WHEN:

AP News - DeepSeek's V4 models were released on April 22nd, 2026.

South China Morning Post - The V4 models were released on April 22nd, 2026.

South China Morning Post - DeepSeek's last model release was V3.2 in December 2025.

WHY:

AP News - Analysts noted that the release of V4 is a "pivotal milestone for China’s AI industry," especially as global competition intensifies in the pursuit of self-reliance in critical technologies.

South China Morning Post - Analysts from Huatai Securities noted that the release of V4 explicitly mentions compatibility with domestic chips, indicating potential improvements in domestic graphics card capabilities.

MarkTechPost - Analysts noted that the release of V4 explicitly mentions compatibility with domestic chips, indicating potential improvements in domestic graphics card capabilities.

South China Morning Post - Analysts from Huatai Securities noted that the release of V4 indicates potential improvements in domestic graphics card capabilities.

Engadget - DeepSeek's V4 models were released following the company's previous app being banned for use by US federal agencies due to national security concerns.

South China Morning Post - DeepSeek claims its V4 model is competitive with leading US closed-source models from OpenAI and Google DeepMind.

South China Morning Post - Analysts from Huatai Securities noted the release of V4 explicitly mentions compatibility with domestic chips, indicating potential improvements in domestic graphics card capabilities.

HOW:

MarkTechPost - DeepSeek's V4 models utilize a Hybrid Attention Architecture that combines Compressed Sparse Attention and Heavily Compressed Attention to achieve a one-million-token context window.

Engadget - DeepSeek's V4 models are still open-source, allowing users to download and modify the code.

Engadget - DeepSeek's V4 models are open-source, allowing users to download and modify the code, and the company claims the models have enhanced agentic capabilities.

Engadget - DeepSeek's V4 models are open-source, allowing users to download and modify the code.

South China Morning Post - Both V4 models have a context window of 1 million tokens, a significant increase from the previous model's context window of 128,000 tokens.

South China Morning Post - The extended technical report outlining V4’s model architecture and training techniques is expected to benefit global AI developers.

(UPDATED April 24, 2026 23:10 UTC)

DeepSeek's new V4 Pro Max version, released on Wednesday, has demonstrated superior performance in standard reasoning benchmarks compared to OpenAI's GPT-5.2 and Google's Gemini 3.0-Pro, although it falls slightly short of GPT-5.4 and Gemini 3.1-Pro. The V4 Pro version also shows promise, potentially outperforming Claude's Sonnet 4.5 and nearing the capabilities of Claude's Opus 4.5 model, according to DeepSeek's evaluations. This release is seen as a pivotal milestone for China's AI industry, especially amid the growing global competition for self-reliance in critical technologies, with the V4 models partly supported by Huawei-made chips that lessen dependence on U.S. manufacturers like Nvidia.

(UPDATED April 24, 2026 21:13 UTC)

DeepSeek has unveiled its V4 Pro and V4 Flash models, with the Pro priced at $1.74 per million input tokens and $3.48 per million output tokens, while the Flash costs just $0.14 and $0.28 for the same amounts. The V4 Pro boasts an impressive 1.6 trillion parameters and a context window of one million tokens, a significant leap from the previous model's 128,000 tokens, and analysts suggest that this release may enhance domestic graphics card capabilities due to its compatibility with local chips. Both models leverage a Hybrid Attention Architecture that effectively combines Compressed Sparse Attention and Heavily Compressed Attention to achieve these advancements.

(UPDATED April 24, 2026 19:17 UTC)

DeepSeek has launched its new V4 models, with DeepSeek-V4-Pro priced at about $5.22 for a million-input, million-output comparison, while the more affordable DeepSeek-V4-Flash costs just $0.42. The V4-Pro model features 49 billion activated parameters per token and is significantly cheaper than competitors, costing roughly one-seventh of GPT-5.5 and one-sixth of Claude Opus 4.7, while the Flash version is over 98% less expensive in a simple input-plus-output comparison. Analysts have noted that these models are designed to work with domestic chips, potentially enhancing graphics card capabilities, and they utilize a Hybrid Attention Architecture for a one-million-token context window.

(UPDATED April 24, 2026 17:18 UTC)

DeepSeek has announced that its V4 models boast "world-class" reasoning capabilities and are described as "cost-effective" with a context length of 1 million tokens. The company claims these models have "closed the gap" with leading competitors on reasoning benchmarks, and analysts from Huatai Securities suggest that their release could signal advancements in domestic graphics card technology. Additionally, the V4 models remain open-source, giving users the ability to download and modify the code as they see fit.

(UPDATED April 24, 2026 15:19 UTC)

DeepSeek has unveiled its V4 Pro model, boasting an impressive 1.6 trillion parameters, while the V4 Flash model features 284 billion parameters. Both models have achieved a remarkable context window of 1 million tokens, a significant leap from the previous model's 128,000 tokens, and are said to be more efficient than the V3.2 version thanks to architectural improvements. Additionally, analysts note that the V4 models' compatibility with domestic chips could enhance graphics card capabilities, and being open-source allows users to download and modify the code for improved agentic capabilities.

(UPDATED April 24, 2026 13:20 UTC)

DeepSeek has announced that its V4 Pro model boasts enhanced agentic capabilities, positioning it as a strong competitor to leading closed-source models in reasoning, while trailing only Gemini-3.1-Pro in terms of rich world knowledge. The company also introduced the V4 Flash model, which offers faster response times and matches the Pro version's performance on simpler tasks. These V4 models come in the wake of DeepSeek's previous app being banned for use by US federal agencies due to national security concerns, and they are open-source, enabling users to download and modify the code.

(UPDATED April 24, 2026 11:18 UTC)

DeepSeek, a Hangzhou-based AI start-up, released its new V4 models on Wednesday, which include the V4 Pro and V4 Flash, boasting 1.6 trillion and 284 billion parameters, respectively. Huawei has announced full support for these models, while Cambricon Technologies confirmed compatibility, highlighting improvements in domestic graphics card capabilities. The V4 Pro model is priced at $1.74 per million input tokens and $3.48 per million output tokens, whereas the V4 Flash model is significantly cheaper at $0.14 and $0.28 per million tokens, respectively.